Vehicle Tracking and Classification From Videos Under Illumination Changes and Occlusions

Vehicle tracking and classification algorithms that remain robust under illuminations changes and occlusions remain a challenging task for vehicle recognition systems. A vehicle which reappears in the scene after disappearing behind an obstacle or a bigger vehicle has to re-obtain the previous identification number assigned by the system. In other circumstances, two or more vehicles overlapping each other are recognized by the system as a single entity: for this reason, after splitting, the system has to reassign pending identification numbers to the respective vehicles. In this paper we propose a three steps (vehicle identification with removal of headlight reflections, tracking with occlusion management and classification with size and speed estimation) algorithm operating in presence of illumination changes, reflections and occlusions. The experimental results obtained by processing a video recorded from a static camera show that the approach is able to successfully manage occlusions in over 90% of cases and to satisfactory classify vehicles into four classes, depending on their length/dimension.